Android Bench is a new open-sourced evaluation platform unveiled at Google I/O 2026 that benchmarks large language models specifically against Android development tasks, including complex workflows such as Jetpack Compose migrations. By focusing on Android-specific scenarios rather than generic coding benchmarks, it gives teams a more relevant signal when selecting models for Android tooling or on-device AI features. The leaderboard is publicly accessible, allowing the wider community to submit and compare model performance. The open-source nature means third-party model providers can run evaluations independently and contribute results, increasing transparency. For Android developers building AI-assisted tooling or evaluating which model to embed in their workflows, Android Bench offers a domain-specific baseline that generic benchmarks like HumanEval or SWE-bench do not provide. The companion Android Skills library, also open-sourced, covers the complex task definitions that drive the benchmark.